Seminares from Emmett Strickland (MoDyCo) and Marc Evrard (LISN)

 

Date: 20/10/2023
Time: 11h00
Localization: IC2, Boardroom
Speakers: Emmett Strickland and Marc Evrard
 
 

Prosody of Nigerian pidgin: corpus approach and experimental validation with a text-to-speech system

 

Résumé: Nigerian pidgin is a poorly endowed language spoken by some 100 million people in West Africa. This talk will present two tools currently being developed to facilitate the study of the relationship between prosody and syntax in this language. The first is a corpus of syntactic trees containing a set of prosodic information on each token. The second is a TTS system that will enable direct control of pitch on synthesised utterances, thus facilitating research into the role of tone and prosody in various perceptual experiments, in particular to validate hypotheses extracted from the corpus.

 
 

Annotation and proposed alternative to annotation for automatic detection of expressive speech

 

Résumé: Work on the annotation of expressive speech in a corpus of political interviews will be presented. The annotation of expressive dimensions such as activation and valence is problematic, particularly for spontaneous or ‘semi-spontaneous’ speech corpora such as political interviews. We propose to discuss an alternative solution for the approximate measurement of activation that does not require annotation. Activation is strongly linked to vocal effort. The aim is to model the estimation of vocal effort using neural architectures trained on calibrated recordings.